Liquid AI releases LFM3-VL-2B, a 1.92B on-device vision-language model with video input and thinking

Liquid AI has released LFM3-VL-2B, a 1.92B-parameter vision-language model for on-device visual understanding. It accepts text, images and video, and is the first LFM vision model that takes video as input, according to the Hugging Face model card.

What's new

The model pairs the LFM3 language model with a new in-house vision encoder. Liquid AI lists three additions:

• Video understanding: questions about recorded lectures, phone walkthroughs or camera-feed clips.
• Thinking: the model reasons step by step in a <think>...</think> block before answering. Liquid AI says this helps most with harder visual questions such as math, charts and multi-step problems.
• Spatial and embodied reasoning: the model locates objects and reasons about their relative placement, which Liquid AI describes as a foundation for robotics applications.

Specifications

Property	Value
Parameters	1.92B (1.73B language model, 164M vision encoder, 25M projector)
Context length	128K tokens (input + output)
Vocabulary	128,000 tokens
Video sampling	2 fps, 4 to 768 frames by default
License	LFM Open License v1.0
Pricing	Not yet disclosed

The language model has 24 layers: 13 gated short convolutions, 5 sliding-window attention and 6 global attention. The vision encoder has 14 layers: 8 gated depthwise 2D convolutions and 6 bidirectional attention layers, using 16×16 patches with 2D rotary embeddings. Inputs are processed at native resolution, and each 2×2 block of patches becomes one language token.

Supported languages are English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian and Polish. The LFM3 family includes base, text, vision, audio, omni and decision models.

Intended use and limits

Liquid AI designs the model for on-device tasks: video questions, reading documents, charts and app screens, finding objects in images, and spatial reasoning for robots. It does not recommend the model for coding, multi-turn conversations or complex reasoning. No benchmark scores are published on the page.

Deployment

Four checkpoint variants are available:

• BF16 for vLLM, SGLang and Transformers (requires transformers 5.21.0 or newer).
• GGUF for llama.cpp and LM Studio.
• MLX with 4-bit quantization for Apple Silicon, via MLX-VLM.
• ONNX for ONNX Runtime.

For most on-device use, Liquid AI recommends llama.cpp b11960 or newer with the GGUF checkpoints. Recommended sampling settings are temperature 0.1, top_p 1.0, top_k 50, min_p 0.0 and repetition_penalty 1.1. In llama.cpp, video always uses 2 fps and 4 to 768 frames, and the --video-fps flag has no effect. Video input requires ffmpeg in the PATH. In Transformers and vLLM, frame rate and frame limit can be lowered to reduce context and memory use.

GPU serving support is not complete. SGLang support is in review (sgl-project/sglang#46858), and vLLM support is in progress. Liquid AI says it will update the card when releases support the model.

LFM3's tool-call format has changed from the custom LFM2/LFM2.5 format to a superset of qwen3_coder, using XML-structured calls.

What this means

At 1.92B parameters, with 4-bit MLX and GGUF builds, LFM3-VL-2B targets local video and image understanding on phones, laptops and embedded hardware rather than server deployment. The 128K context window is shared across input and output, and Liquid AI notes that each video frame adds vision tokens, so long clips need a lower frame rate or frame cap.

Teams moving from LFM2 or LFM2.5 need to update tool-calling code for the new format. Anyone planning high-throughput GPU serving should wait for vLLM and SGLang releases that support the architecture.